Temporal network metrics and their application to real world networks John Kit Tang Robinson College University of Cambridge 2011 This dissertation is submitted for the degree of Doctor of Philosophy Declaration This dissertation is the result of my own work and includes nothing which is the outcome of work done in collaboration except where specifically indicated in the text. This dissertation does not exceed the regulation length of 60 000 words, including tables and footnotes. Summary The analysis of real social, biological and technological networks has attracted a lot of attention as technological advances have given us a wealth of empirical data. Classic studies looked at analysing static or aggregated networks, i.e., networks that do not change over time or built as the results of aggregation of information over a certain period of time. Given the soaring collections of measurements related to very large, real network traces, researchers are quickly starting to realise that connections are inherently varying over time and exhibit more dimensionality than static analysis can capture. This motivates the work in this dissertation: new tools for temporal complex network analysis are required when analysing real networks that inherently change over time. Firstly, we introduce the temporal graph model and formalise the notion of shortest temporal paths, used extensively in graph theory, and show that as static graphs ignore the time order of contacts, the available links are overestimated and the true shortest paths are underestimated. In addition, contrary to intuition, we find that slowly evolving graphs can be efficient for information dissemination due to small-world behaviour in temporal graphs. Secondly, we then turn our attention to the identification of important or central nodes in a network. Since two key mea- sures for node centrality, namely closeness and betweenness, are based on shortest paths in a static graph, we define temporal centrality based on temporal shortest paths. We demonstrate that the ranking achieved by temporal centrality is supe- rior to static analysis by demonstrating how temporal centrality can be exploited to improve mobile malware containment. Thirdly, we study the predictability of cen- trality ranking in temporal networks utilising correlogram plots between top-k node rankings. We show that in real human contact networks, temporal centrality can be predicted and demonstrate that these predictions are useful for mobile malware containment, compared to static centrality prediction. Finally, we investigate the concepts of temporally connected components and show that temporal analysis gives us a precise understanding of the diffusion properties of real contact networks that is missed by static analysis. The conclusions of this thesis are that the use of time aware metrics for the analysis of real networks opens the doors to more precise and effective exploitation of complex network science: while we have given a number of application examples, the future directions of this research are still many. Acknowledgments Firstly, I am indebted to my supervisor and teacher, Cecilia Mascolo, for her con- tinual support, guidance and nurture to become an independent researcher; looking back through the years of electronic correspondence I now realise how much I have learnt from her and the multitude of collaborations that she has opened up for me. Secondly, I thank Mirco Musolesi who played a huge part in mentoring me through the precise process of conducting research and scientific writing. Thirdly, I thank Vito Latora and Murtaza Zafer who both taught me incredible lessons which shaped my own research philosophy; Hyoungshick Kim, Vincenzo Nicosia and Salva- tore Scellato for invigorating collaboration; and Jon Crowcroft and Ross Anderson for their wisdom on guiding my thesis. Also, I thank the friends I have made in the lab, especially Salvo, Kiran, Liam, Bence, Ilias, Christos and Tassos for making my life at Cambridge more enjoyable. I thank the EPSRC and Lise Gough for providing me the financial support which enabled me to complete this PhD; Robinson College and the Cambridge Philosophi- cal Society for generous travel grants; Piete Brooks for rearing the Condor Compute Grid; IBM Research for the stimulating work experience; and Philip Treleaven for giving me a taste for research as an undergraduate and for the opportunities he has opened up for me. I thank my parents and grandparents for their sacrifices in supporting me through my education and their wisdom which has shaped the person I am today. I thank my brother, Eric, and sister, Katie, whose successes have inspired me. I thank my friends (Lovejoy, Oscar, Asbjørn,˚ Jenny, Janet, Mohsen, Attard and many others) for keeping me sane; and I thank Carole, Charlotte, Sylvia and Graham for their loving support and Sunday roasts. Finally, this dissertation is dedicated to my loving Wife, Annabel; without her endless love, patience and motivation I would never have been able to complete this journey: Amor Vincit Omnia. In loving memory of Grandma. Contents Declaration i Summary ii Acknowledgments iii Contents v 1 Introduction 1 1.1 Real Networks Change Over Time . .3 1.1.1 Social Networks . .4 1.1.2 Biological Networks . 10 1.1.3 Technological Networks . 12 1.1.4 Urban Networks . 17 1.1.5 Summary and Discussion . 18 1.2 Contributions . 21 1.3 Chapter Outline . 23 1.4 List of Publications . 23 2 Static Complex Network Theory 25 2.1 Static Model . 26 CONTENTS CONTENTS 2.2 Static Analysis . 27 2.2.1 Small-world metrics . 27 2.2.2 Efficiency . 30 2.2.3 Centrality . 31 2.2.4 Reachability . 33 2.3 Conclusions . 36 3 Temporal Graphs and Distance Metrics 37 3.1 Temporal Graphs . 38 3.1.1 Simplifying Assumption . 40 3.2 Temporal Metrics . 41 3.2.1 Temporal paths and shortest path length . 41 3.2.2 Example calculation of dij .................... 42 3.2.3 Algorithm & Complexity . 45 3.2.4 Temporal distance is a quasi-metric . 52 3.2.5 Characteristic Temporal Path Length . 52 3.2.6 Local Temporal Efficiency . 53 3.2.7 Temporal Correlation Coefficient . 53 3.3 Literature Review . 54 3.3.1 Introduction . 54 3.3.2 Related Work . 54 3.3.3 Discussion . 59 3.4 Application to Real Networks . 61 3.4.1 Introduction . 61 3.4.2 Importance of Time in Real Networks . 62 3.4.3 Small-world Behaviour in Temporal Graphs . 71 3.5 Conclusions . 78 vi CONTENTS CONTENTS 4 Temporal Centrality Measures 80 4.1 Temporal Centrality . 82 4.1.1 Temporal Betweenness Centrality . 82 4.1.2 Temporal Closeness Centrality . 83 4.1.3 Runtime Complexity . 84 4.2 Application to Real Networks . 84 4.2.1 Corporate Email Dataset . 84 4.2.2 Short Range Mobile Malware Containment . 92 4.3 Related work . 108 4.4 Conclusions . 109 5 Predicting Information Spreaders in Temporal Graphs 111 5.1 Top-k Prediction Model . 113 5.1.1 Example . 113 5.1.2 Parameters . 115 5.2 Predictability of Human Contact Traces . 116 5.2.1 Top-k Correlation Function . 116 5.2.2 Testing for Top-k Correlations . 117 5.2.3 Prediction Function Design . 118 5.3 Application to Real Networks . 119 5.3.1 Parameters and Evaluation Metrics . 119 5.3.2 Effect of Malware Start Time . 120 5.3.3 Increasing Patch Delay . 124 5.3.4 Effects of Contact Upload Interval . 124 5.3.5 Varying initial compromised and patched devices . 124 5.4 Related work . 126 5.5 Conclusions . 127 vii CONTENTS CONTENTS 6 Reachability in Temporal Graphs 129 6.1 Temporally Connected Components . 130 6.2 The affine graph of a temporal graph . 133 6.3 Application to a Real Network . 137 6.4 Related Work . 145 6.5 Conclusions . 146 7 Summary and Outlook 148 Bibliography 150 viii 1 Introduction Networks are all around us: from the cities and roads that we live in to the physi- cal telecommunication cables that connect our computers forming the Internet, and from the intricate layout of neurons and synapses that drive our brains to the re- lationships between friends; the term \networks", whether road, computer, online social or otherwise, has now become common in our everyday vocabulary. Though the term has been integrated into our culture, the analysis of such a topological ab- straction is in itself still a growing science where everyday networks can be modelled as a set of nodes (e.g., cities, computers, neurons or people) which are connected by edges (e.g., roads, cables, synapses or relationships) and the analysis of the non- trivial features of such networks has opened a branch of study known as complex network analysis [AB02]. At its roots, complex network analysis is founded on graph theory [BLM+06] and hence networks are commonly referred to as graphs. Indeed the publication regarded as the beginnings of graph theory was that of Leon- hard Euler's study on the seven bridges of K¨onigsberg, published in 1736, which posed the question of whether a walk existed through the city of K¨onigsbergs, which 1 2 is divided into four landmasses by its river, that would cross its seven bridges once and only once. By mapping the city into a topological graph representation with landmasses as nodes and bridges as edges (depicted visually in Figure 1.1), Euler was able to reason on this graph and prove that there was in fact no solution. 1 3 1 4 2 3 4 2 (a) Map (b) Topological Figure 1.1: Example of topological mapping of the seven bridges of K¨onigsbergs to a graph. K¨onigsbergs was split by its river into four land masses. The graph in panel (b) is visually depicted as nodes (circles) and edges (lines connecting circles).